This paper presents the application of an ellipsoidal method for robust dynamic identification of parallel robots. The robot is modelled with classical Lagrange equation which leads to an inverse dynamic model linear with respect to the parameters. Assuming the error additive on input (motor torque), the problem is expressed in a bounded error context. The ellipsoidal method is applied in a factorised form in order to guarantee numerical stability. Experimental results are exhibited for a fully parallel robot with 4 degrees of freedom.